What's this about?
Research showing that AI-agent safety failures often originate in how tools are specified — the descriptions and schemas developers write — not just in model behavior, and proposing mitigations at the specification layer.
What exactly is a "tool specification"?
When you give an agent a tool (send email, run code), you describe it in text: what it does, its parameters, when to use it. The model acts based on that description. A sloppy or ambiguous spec is like a mislabeled control panel — the operator may be competent, but the wrong label still gets the wrong button pressed.
Why it matters
It shifts some safety burden from "align the model" to "audit your tool definitions" — a much more actionable checklist for anyone shipping agents today.